transformer-circuits

A Mechanistic Interpretability Toolkit for Cross-Layer Transcoder Training and Attribution-Graph Visualization

The Dataset and Official Implementation for <Discursive Circuits: How Do Language Models Understand Discourse Relations?> @ EMNLP 2025

Tracking exactly what happens to the internal "circuitry" (induction heads) of a 2-layer attention-only Transformer when forced to undergo domain adaptation from prose to structured Python code.

Mechanistic interpretability study: identifying the 4 attention heads (out of 144) that control 83% of subject-verb agreement in GPT-2 Small

Open-source EU AI Act Annex IV documentation toolkit. Mechanistic interpretability + circuit discovery for transformers. One function call generates a structured, hash-chained evidence package.

A Jacobian-Lens (J-Lens) observer for vision-language models — read what a VLM is poised to say, before it says it. Multimodal J-Lens on Qwen3.5, concept-race, and a forward-only prompt helper.

Does Quantization Kill Interpretability? Scaling study across 5 models (124M-2.8B): RTN destroys induction heads in small models, GPTQ preserves them at all scales.

Reproduction of the induction-heads circuit (Olsson et al., 2022) in a 2-layer attention-only transformer trained on synthetic random-boundary repeat sequences, with mechanistic identification by ablation and training-time tracking of circuit formation.

Reverse-engineering neural network internals from scratch in NumPy + PyTorch. A 6-week masterclass: linear representation hypothesis, superposition, sparse autoencoders, transformer circuits & induction heads, activation/path patching & causal scrubbing, and steering a real LM. Fully executed notebooks.

Chain-of-density study of 'Emotion Concepts and their Function in a Large Language Model' (Sofroniew et al. 2026, Transformer Circuits/Anthropic, arXiv:2604.07729) - five-tier note, locator-verified. Emotion vectors, steering, functional emotions, desperation-blackmail dial.

Natural Language Autoencoder (NLA) research prototype inspired by Anthropic’s interpretability work. Implements a scoped approximation of activation verbalization and reconstruction on small open-source LLMs, with quantitative evaluation, baselines, and reproducible local-first experimentation.

Chain-of-density study of 'Verbalizable Representations Form a Global Workspace in Language Models' (Gurnee, Sofroniew, Lindsey et al. 2026, Transformer Circuits/Anthropic) - five-tier note, locator-verified. The Jacobian lens, the J-space, alignment auditing, counterfactual reflection training.

A systematic review, learning path, and reproducible research map for AI safety, mechanistic interpretability, and transformer circuits

Visualize and manipulate GGUF models using Jacobian Lens steering, concept patching, and ablation for llama.cpp.